Debiasing Counterfactuals In the Presence of Spurious Correlations

• DOI: 10.48550/arxiv.2308.10984
• Publication Date: 2023-01-01
• Authors: Amar Kumar; Nima Fathi; Raghav Mehta; Brennan Nichyporuk; Jean-Pierre Falet; Sotirios A. Tsaft...; Tal Arbel
• Keywords: Spurious relationship; Debiasing

Deep learning models can perform well in complex medical imaging classification tasks, even when basing their conclusions on spurious correlations (i.e. confounders), should they be prevalent the training dataset, rather than causal image markers of interest. This would thereby limit ability to generalize across population. Explainability based counterfactual generation used expose confounders but does not provide a strategy mitigate bias. In this work, we introduce first end-to-end framework that integrates both (i) popular debiasing classifiers (e.g. distributionally robust optimization (DRO)) avoid latching onto and (ii) unveil generalizable relevance task. Additionally, propose novel metric, Spurious Correlation Latching Score (SCLS), quantify extent classifier reliance correlation as exposed by images. Through comprehensive experiments two public datasets (with simulated real visual artifacts), demonstrate method: learns population, successfully ignores focuses underlying disease pathology.